Skip to content

OpenAI's GPT-6.2 Astra and Sol ship with a breaking change to reasoning caching

· by

The short version

OpenAI's November 17 changelog introduces GPT-6.2 Astra and Sol and a Responses API change that drops saved reasoning after a conversation edit, plus new discounted background completion windows.

OpenAI's GPT-6.2 Astra and Sol ship with a breaking change to reasoning caching

OpenAI shipped two new models, GPT-6.2 Astra and GPT-6.2 Sol, on November 17, and the migration notes carry a detail most changelogs bury: edit an earlier message or tool call in a conversation and the Responses API now discards the model's saved reasoning, per OpenAI's migration guide, which can lower answer quality and burn extra tokens as the model redoes work it had already finished. The same release also moved where a reusable system prompt needs to live if a team wants it to keep hitting the prompt cache.

What happened

OpenAI's API changelog confirmed GPT-6.2 Astra, built for demanding reasoning and coding tasks, and GPT-6.2 Sol, aimed at everyday coding and agent workflows, both went live November 17, per OpenAI's API changelog. Both models work through the Responses API and Chat Completions, but tool calling only works through the Responses API, the changelog states. GitHub made both models generally available the same day inside Copilot, limited to Pro+, Max, Business, and Enterprise plans, not the base Copilot Pro tier, per GitHub's changelog. In GitHub's own early testing, GPT-6.2 Sol caught nearly twice as many real code-review issues as GPT-6.1 Sol, and GPT-6.2 Astra caught more issues than GPT-6 Astra while using fewer tokens on agentic coding tasks, according to that same changelog post. GPT-6.2 Sol carries the exact API pricing of its predecessor: $2.00 per million input tokens, $0.10 for cached input, $2.50 for a 30-minute cache write, and $10.00 per million output tokens, per OpenAI's GPT-6.2 Sol model page, with those rates holding up to 272,000 input tokens before input and cache pricing double.

Reusable system prompts need a new home to stay cached

Two Responses API behaviors change starting with GPT-6.2, and both can bite integrations that already exist, per OpenAI's migration guide. First, the API now drops a conversation's saved reasoning after what OpenAI calls a prefix edit: a change to an earlier developer message or tool call, a model downgrade, or a change to the end user or organization tied to the request. Losing that reasoning can lower answer quality and push the model to redo earlier work, which spends more tokens than before. Second, GPT-6.2 models append top-level instructions after the input array rather than before it, which breaks the exact-prefix match that prompt caching depends on. OpenAI's own fix is to move any reusable system prompt into a developer message placed first in the input array, and to use the top-level instructions field only for guidance specific to a single request. Any app that lets a user or an earlier tool call get edited mid-conversation, or that leans on a static top-level instructions block for caching, needs to check both behaviors before routing production traffic to GPT-6.2.

The same November 17 changelog batch also added a cost lever unrelated to the new models themselves: a public-beta choice of a two-hour or eight-hour background completion window in the Responses API, with token rates 30% lower than standard for the two-hour window and 40% lower for the eight-hour window, per OpenAI's changelog. The discount applies to GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, and GPT-6 Luna, so any async or batch-style workload already calling the Responses API in the background can pick up the lower rate without switching models.

Context

GPT-6.2 Sol succeeds GPT-6.1 Sol, which OpenAI released September 29, 2026 at the same $2 input and $10 output pricing, with Multi-agent delegation in beta, per GitHub's changelog from that rollout. GPT-6 Astra, the model GPT-6.2 Astra replaces at the top of the lineup, reached general availability in GitHub Copilot on September 4, 2026, per GitHub's changelog. OpenAI's deprecations page currently lists GPT-6 Astra and GPT-6.1 Sol only as the recommended replacement models for the outgoing GPT-5 family, which is scheduled for deprecation by December 11, 2026, per OpenAI's deprecations page; as of this writing, OpenAI has not posted a retirement date for GPT-6 Astra or GPT-6.1 Sol themselves.

What to watch next

Whether OpenAI sets a deprecation date for GPT-6 Astra and GPT-6.1 Sol now that their successors have shipped is the open question; the deprecations page has not added one yet. The two-hour and eight-hour background completion discounts remain in public beta, and OpenAI has not said when, or whether, that pricing becomes the standard rate for background requests.

Sources